v1.8.0

Added

• MCP over Streamable HTTP at /mcp on the REST server — many sessions, one store. (#92)

  Until now the only MCP transport was stdio, so every Claude Code session spawned its own server, and two sessions on one repo (a main checkout and a worktree) put two processes on one SQLite file, which docs/concurrency.md forbids. A hosted instance (Railway, Cloud Run, the Docker image) was reachable by REST only, which no MCP client speaks. fozikio serve --rest now also serves MCP at /mcp (src/mcp/http.ts): stateful sessions keyed by the SDK's mcp-session-id, one Server + StreamableHTTPServerTransport per client over the one EngineContext, the same tool list as stdio (the /api/tools/:name blocklist guards the generic REST endpoint, not an agent's own session). Auth is the REST token, in x-cortex-token or Authorization: Bearer. An unknown session id answers 404 and a non-initialize request with no session 400, as the spec says. .mcp.json for a local shared server or a hosted one: { "type": "http", "url": "…/mcp", "headers": { "x-cortex-token": "${CORTEX_API_TOKEN}" } }. createMcpServer(engine) builds a Server over an existing engine; createServer(config) wraps it as before. startRestServer returns the http.Server (it returned void) so a caller can read the bound port or close it; the start-up log names the /mcp URL. Tests drive the mounted endpoint with the SDK's own client: initialize → tools/list → tools/call, both auth headers and the 401, two clients with distinct sessions over one engine, the 404/400 cases. docs/deploy-railway.md, docker-compose.yml, the README and docs/concurrency.md say how to use it.

  Observed while proving it against a live store: the first LLM-backed call after ollama loads a 14B model cold took about 80 s, past the SDK client's 60 s default request timeout; warm calls take seconds. A client that may hit a cold model should raise its per-request timeout, or the model should be kept loaded.